The Bayesian trial design conversation tends to start after patients are assigned to arms. Priors, borrowing, monitoring, posterior inference. All of it assumes the randomization is settled. Stratified permuted block, 1:1 allocation, done. But how you randomize affects power, balance, ethical allocation, and regulatory credibility. Most biostatisticians treat it
My public comments on Docket No. FDA-2025-D-3217, "Use of Bayesian Methodology in Clinical Trials of Drug and Biological Products."
When the Pfizer/BioNTech BNT162b2 trial reported 95% efficacy in November 2020, the world saw a scientific triumph. What most people missed, and what many statisticians still underappreciate, is that the trial's primary analysis was Bayesian. Not frequentist group sequential boundaries. Not O'Brien-Fleming. A posterior probability
In January 2026, the FDA released updated guidance on the use of Bayesian methods in clinical trials. The document does not read like a manifesto. It reads like an attempt to reconcile competing statistical cultures under real regulatory constraints. On its surface, it is pragmatic and flexible, welcoming Bayesian designs
In 2019, FDA granted accelerated approval to voxelotor for sickle cell disease based on a surrogate endpoint: hemoglobin increase. The mechanism was sound. Voxelotor increased hemoglobin oxygen affinity, which should reduce sickling and improve outcomes. The biology made sense. The FDA agreed. Patients got access. In September 2024, the drug
Most statistical frameworks treat errors symmetrically. A false positive is bad. A false negative is bad. Control one, tolerate the other, and let the math do the rest. Clinical reality is not that tidy. Approving an ineffective therapy and withholding a potentially effective one are both errors, but they do
In September 2016, the FDA approved eteplirsen for Duchenne muscular dystrophy. The advisory committee had voted 7 to 6 against accelerated approval. The FDA's own review team recommended against it. The clinical program consisted of 12 boys, and western blot analysis showed a dystrophin increase of 0.93%
The Bayesian vs. frequentist debate usually centers on philosophy. Priors are subjective! No, they formalize existing knowledge! You're smuggling in assumptions! You're ignoring relevant information! It's a fun argument. It's also the wrong one, at least for regulatory decision-making. The real value
Between late 2024 and late 2025, FDA approved six rare-disease therapies supported primarily by single-arm trials. None of these sponsors ran randomized controlled trials. All received traditional or accelerated approval. What separated success from rejection wasn’t luck or regulatory leniency. It was understanding what evidence compensates for the absence
I asked an LLM a question that junior biostatisticians ask senior biostatisticians all the time: Should I use a Bayesian borrowing design for my Phase 2 single-arm oncology trial? Here's what I got back. The LLM Response Should I use a Bayesian borrowing design for my Phase 2
What FDA Actually Decided When the Trial Failed
If you’ve ever justified a Bayesian design to a regulator, you’ve already been practicing calibrated Bayes, even if you never called it that.
What REBYOTA teaches us about dynamic borrowing
The FDA just gave us 25 pages on Bayesian inference in pivotal trials. The philosophical message is clear: Bayesian methods are welcome. Informative priors, borrowing from external data, direct interpretation of posterior probabilities were all explicitly endorsed. The operational message is different: pre-specify everything. Bayesian methods promise learning. Regulators demand
I’ve been in meetings where “adaptive” was treated as a synonym for “smaller trial.” The assumption goes like this: if we can learn as we go, we can stop early when something works, drop arms that fail, and route patients to better treatments. Surely that means fewer patients overall.
The pharmaceutical industry is facing a cost disease. Bringing a new drug to market now routinely exceeds $2 billion, with patient recruitment standing as the single largest bottleneck. In response, we usually reach for operational fixes: faster sites, simpler protocols, digital recruitment. But there is a mathematical lever that is
What to do instead of UPSIs
The most expensive decision in drug development (or product experimentation) is not starting a trial. It is continuing a zombie trial. Imagine you are at the interim analysis of a Phase II study. You have enrolled 50 of 100 patients. You have spent $10 million. The data is... underwhelming. The
Everyone says they want adaptive designs. Almost no one actually runs them. In decks, protocols, and FDA briefing books, “adaptive” has become a fashionable adjective—usually meaning a trial with one interim look, a conditional power calculation, and a long list of things you are not allowed to change. The
In the late 1980s, cardiologists thought they had cracked the code. They knew that patients who developed irregular heartbeats, premature ventricular contractions (PVCs), after a heart attack were more likely to die. They had drugs, encainide and flecainide, that reliably suppressed those irregular beats. The intuition was flawless: Suppress the
What oncology learned the hard way—and tech keeps relearning.
In 2014, during a major snowstorm, Uber ran a randomized test by turning off surge pricing for a subset of riders. Some waited longer. Others never got a ride at all. It wasn’t just a pricing tweak. It was a trial of behavior, supply, and equity. Trials, once reserved
Lessons from EDI and Zelen’s Design Dilemma
Rethinking Trial Design Through ISIS-2
No posts found matching your search.
